Multiphase Contrast-Enhanced CT based Prediction of VETC/MVI Phenotypes in Hepatocellular Carcinoma

Abstract

Background: Hepatocellular carcinoma (HCC) exhibits substantial biological heterogeneity, and conventional imaging assessment is limited in its ability to characterize aggressive vascular phenotypes before surgery. Among these phenotypes, microvascular invasion (MVI) and vessels encapsulating tumor clusters (VETC) represent related but biologically distinct markers of tumor aggressiveness. Integrating MVI and VETC into a unified four-class phenotype framework may provide a more comprehensive preoperative characterization of HCC than predicting either feature alone. Multiphase contrast-enhanced CT is well suited to this task because it captures phase-dependent enhancement patterns and tumor heterogeneity across arterial, portal venous, and delayed phases.Purpose: To develop and comparatively evaluate triphasic contrast-enhanced CT-based quantitative imaging approaches for preoperative four-class prediction of combined VETC/MVI phenotypes in hepatocellular carcinoma, including single-phase radiomics multiphase feature fusion, delta radiomics, deep feature fusion, and habitat-based heterogeneity modeling, while additionally assessing standalone deep learning as a supplementary comparator.Materials and Methods: This retrospective single-center imaging biomarker study included 366 patients with pathologically confirmed HCC who underwent triphasic contrast-enhanced CT between May 2013 and July 2023. The cohort comprised 148 patients with VETC-/MVI-, 82 with VETC+/MVI+, 61 with VETC-/MVI+, and 75 with VETC+/MVI-. Tumors were manually segmented on AP, PVP, and DP images. Six primary quantitative imaging pipelines were constructed: single-phase radiomics, multiphase radiomics fusion, delta radiomics, deep feature fusion, habitat-only features, and habitat radiomics. Model performance was evaluated using macro-average one-vs-rest area under the receiver operating characteristic curve (macro-AUC), balanced accuracy, macro-F1 score, and Matthews correlation coefficient (MCC).Results: Delta radiomics achieved the highest overall discriminative performance. The best-performing primary model was the DP-PVP XGBoost model, which reached a macro-AUC of 0.847, the highest among all evaluated primary pipelines. Habitat radiomics also performed strongly. Within this branch, the highest macro-AUC was 0.839 for the DP SVM model, while the AP logistic regression model achieved a balanced accuracy of 0.625, a macro-F1 score of 0.614, and an MCC of 0.530. Deep feature fusion improved upon deep-feature-only embeddings but remained inferior to the strongest habitat-radiomics and delta-radiomics models. In contrast, multiphase early fusion and low-dimensional habitat-only features showed weaker performance. Overall, approaches that explicitly modeled heterogeneity across phases or within tumor subregions outperformed simpler whole-tumor and direct-concatenation strategies.Conclusion: Triphasic CT-based quantitative imaging shows promise for noninvasive preoperative prediction of combined VETC/MVI phenotypes in HCC. Among the evaluated approaches, delta radiomics delivered the strongest overall discriminative performance, whereas habitat radiomics showed particularly strong classification-oriented performance and robust branch-level results. These findings suggest that heterogeneity-aware modeling, particularly inter-phase delta radiomics and subregion-aware habitat radiomics, may be especially effective for biologically informed preoperative phenotyping of HCC.